January 21, 2006

Reconciling the Returns to Education in Off-Farm Wage Employment inRural China

Alan de Brauw and Scott Rozelle∗

∗Alan de Brauw is assistant professor, Department of Economics, Williams College, and Scott Rozelle is pro-fessor, Department of Agricultural and Resource Economics, University of California, Davis. Please direct all cor-respondence to Alan de Brauw, Department of Economics, Williams College, Williamstown, MA 01267 (e-mail:alan.d.debrauw@williams.edu; phone: (413)597-4373; FAX: (413)597-4045). We are grateful to Belton Fleisher, Qi-uqiong Huang, Albert Park, J. Edward Taylor, Jim Wilen, Yigang Zhang, and an anonymous referee for comments onthe manuscript. The authors acknowledge the support of the Ford Foundation, Beijing. Rozelle is a member of theGiannini Foundation of Agricultural Economics.

Reconciling the Returns to Education in Off-Farm Wage Employment in

Rural China

Abstract

Previous studies have found that the returns to education in rural China are far lower than

estimates for market economies. In this paper, we try to determine why previous estimates are

so low. Whereas estimates for the early 1990s average 2.3 percent, we find an average return

of 6.4 percent. Furthermore, we find even higher returns among younger people, migrants,

and for post-primary education. We further show that although part of the difference between

our estimate and previous estimates can be attributed to increasing returns over time, a larger

portion of the difference is due to the methodology used by other authors.

1

Reconciling the Returns to Education in Off-Farm Wage Employment in

Rural China

Rural education is of primary importance to development in China. Since family planning

policies were implemented, birth rates have been much higher in rural areas than in urban areas. As

approximately 60 percent of China’s population permanently resides in rural areas, an even larger

percentage of children are educated in rural areas (NBS, 2005). Therefore, the effectiveness of rural

education is of particular interest to policy makers and researchers interested China’s continuing

development.

Despite the importance of building human capital in rural China, education has not had a

prominent role in government development efforts during the past two decades, and faces serious

fiscal problems (West, 1997; Nyberg and Rozelle, 1999). Unlike its successful neighbors in East

Asia, China’s central government has traditionally spent relatively little on rural education. For

example, Heckman (2003) reports that China as a whole spent 2.5% of GDP on education, while

the world average is 5.2%, and other developing countries in Asia typically spend between 4 and

5%. Rural education attainment rates are also relatively low for its development level. In China,

the average educational attainment is only 6.1 years, compared with more than eight years in the

rest of Asia (Psacharopoulos, 1994).

Education may be underfunded in China in part because investments in rural education are

perceived to generate relatively low private, and therefore social, rates of return. While measured

average returns to education in the developing world are over 10 percent and exceed 9 percent in

other Asian countries (Psacharopoulos, 1994), studies in rural China typically find returns below 5

percent. In the urban economy, studies using data collected in the late 1980s and early 1990s rarely

found returns to exceed 5 percent (e.g., Appleton et al., 2002; Meng and Zhang, 2001; Maurer-

Fazio, 1999).1 Studies of the rural economy have typically found even lower returns (e.g., Meng,

1996; Parish, Zhe, and Li, 1995; Zhao, 1999). If previous studies are correct and investment in

2

schooling in China does not lead to higher incomes, it can be argued that the central government

should reduce its emphasis on schooling in its development plan in favor of other expenditures that

will lead to higher returns.

As returns to education are relatively high in other countries, consistently low estimates of

returns in rural China demand further explanation. Given its socialist legacy, it is possible that

China’s reformers may have insulated managers from the pressures of markets. Managers may

have been encouraged to use non-market factors to assign jobs to workers rather than hiring more

educated workers or workers with better qualifications. While this explanation is plausible for

the early years of transition in urban China, it is unlikely to be the major reason for low returns

in the rural economy, as it became market-oriented much earlier in the market transition than the

urban economy. Rural enterprises have operated in an increasingly competitive environment, and

managers in most fields have had fairly good incentives and rights (e.g. Naughton, 1995; Chen and

Rozelle, 1999; Jin and Qian, 1998).

Alternatively, the returns to education in rural China may have been systematically underes-

timated due to methodological shortcomings. Previous studies may have mismeasured wages by

using a wage measure that endogenizes the individual’s decision regarding the amount of labor to

allocate to off-farm work. As a result, wages could have been systematically understated for the

educated relative to the less educated. Using data that allow for a more appropriate measure of the

wage– the hourly wage instead of a daily or monthly wage– could help provide better estimates.

Furthermore, previous studies may have focused on only one sector of the rural off-farm economy.

If that sector– for example, rural industry– was dominated by firms with managers that did not

reward workers for their education, the low measured rates of return may be due to the fact that

the firms were simply not representative of the rural economy as a whole. Finally, returns could

be lower for low levels of education (primary school) relative to higher levels of education (lower

and upper secondary school). Treating each year of education as the same could lead to systematic

underestimates of the returns to education in studies of an economy, like China’s, that has only

3

been successful in reaching an average level of attainment of primary school (Strauss and Thomas,

1995). Regardless of the source, if low estimates of the rate of return to schooling are primarily

due to methodological shortcomings, previous research may have contributed to the problem of

low investment in education rather than helping solve it.

The goal of this paper is to help explain the gap between estimated returns to education in

off-farm employment in rural China and the rest of the world. We produce new estimates of the

returns to education in rural China and reconcile our results with those in the literature. While it is

possible earlier studies find low returns due to institutional constraints, we show that methodology

played a major role in the low measured rates of return that characterize the human capital literature

on rural China. We demonstrate that three factors have helped lead to these low measurements–

mismeasured wage rates, unrepresentative samples, and a failure to account for the non-linearity

of returns to education. We find that the methodological differences between our study and others

can account for a significant part of the differences between rates of return for rural China and

those found in the rest of the world. In fact, several subsets of our estimates are almost completely

consistent with those discussed by Psacharopoulos (1994) for East Asia in general.

To accomplish this goal, the paper will proceed as follows. The next two sections describe

our data set and examine previous work on wages and rates of return in rural China. In the fol-

lowing sections, we describe the empirical framework that we will use to produce a new set of

estimates of rural rates of return and present our baseline estimates. After examining how the es-

timates change when alternative specifications, samples, and estimations techniques are used, we

show how several key assumptions can explain the previous results in the rural China education

literature and how relaxing those assumptions make China’s results come in line with those in the

rest of the world. The final section concludes.

4

1 Data

The data set used in this paper is from a nearly nationally representative survey of 1199 house-

holds in 6 provinces and 60 villages in rural China conducted by the authors in late 2000.2 In

addition to collecting basic information on the farm household, land and labor endowments, and

other production-oriented activities, the survey included detailed information about labor force

participation and schooling among all household members. Enumerators questioned all household

members about their education, and asked all household members about their employment and

work except for children under fifteen, individuals in school, and the elderly who no longer work.

In total, our sample includes 3364 individuals who were 16 years of age or older who could po-

tentially work. Of those, 1022 individuals worked for a wage off the farm either locally or as a

migrant, and 2341 did not.

To assess the representativeness of our sample, we compare employment rates in different

broad categories in our sample with the national statistics for 2000 (NBS, 2001). Some differences

should be expected, as we were able to pick up migrants in our sample and categorize them as

off-farm workers. We also allow workers to enter multiple categories in our sample, as we observe

individuals both working on the farm and in wage earning jobs. Keeping these differences in

definition in mind, we find that our sample is largely consistent with the national statistics (Table 1).

In the national statistics, 67% of rural workers were reported as primarily employed in agriculture.

We find that 76% of workers spent some time working in agriculture in our sample. However, some

of those people worked both on- and off-farm, which would lead us to find higher participation,

and we include migrants, which would lead us to find lower participation. If we remove migrants

from our sample, we find that 36% of the remaining workers either spent some time working for

a wage or in self-employment. While this figure suggests lower primary sector employment than

the national statistics, some of the workers in our sample spent only a small fraction of their time

off-farm, implying that our figures are roughly consistent with the national ones.

5

Several aspects of both the household and village level surveys were designed specifically to

try to help answer the questions raised in this paper. The survey asked about school participation

using both the years a person was in school and the final level of school attainment, and specifically

asked if people repeated grade levels. Hourly wages were computed by taking all monetary earn-

ings over the course of the year (in multiple jobs, if the person held more than one wage earning

job) and dividing by the total number of hours that were reported as being worked during the year.

It also asked respondents to identify if they lived at home while they were working, or if they lived

away from home, so they could be categorized as local wage earners or migrants. Migrants include

individuals who are still considered household members, but they lived outside the household for

at least one month during the year while working. Migrants should still be considered part of the

rural labor force, because they take jobs that urban laborers will not (Rozelle et al., 1999). More-

over, many migrants in our sample work in rural areas; only half of them leave their home county.

To proxy for individual ability, which is known to bias estimates of the returns to schooling, we

follow Ashenfelter and Zimmerman (1997) and use questions enumerators asked about the years

of schooling each person’s parents attained.

2 Returns, Educational Attainment, and Methodological Short-

comings

Meta-analyses of studies of the returns to education show a striking regularity in the relationship

between wages and schooling levels in developing countries (Figure 1).3 In Sub-Saharan Africa

and Latin America, where average educational attainment is between 5 and 8 years, an individual

who attends an additional year of schooling gains a return of more than 12 percent annually in the

form of increased wages. In Asia and the Middle East, where average educational attainment is

higher, rates of return, although somewhat lower, still typically range between 8 and 10 percent

6

(Psacharopoulos, 1994). Average returns in the OECD countries, where most of the population

graduates from high school, still reach nearly 7 percent.

In rural China, the average level of educational attainment, although rising in recent years,

is significantly lower than the level in the rest of Asia (Figure 1). Among the entire workforce, the

average years of schooling is 6.13 years. Among workers 35 years old and younger, the average

education level is much closer to the level in the rest of Asia, at 7.59 years. Education levels have

increased among young people as a result of educational reforms that have, among other changes,

mandated that children go to school for nine years. However, there is no guarantee, especially in

poor areas, that students go to school for the mandated nine years (Tsang, 1996; Brown and Park,

2002).

Rural China’s pattern of educational attainment and returns does not fit the pattern found in

the rest of the world. In rural China, estimates of the return to schooling made by other researchers

are consistently quite low (Table 1). In our search of the literature, we identified six studies of rates

of return in rural China that have used standard Mincerian methods.4 The average return across

these studies is about 4 percent. Only one of the studies, Ho et al. (2001), measures returns in

excess of 5 percent. Plotted on Figure 1, China stands out as having abnormally low rates of return

to education when compared with the rest of the world, given its average educational level.

In addition to the experience of other developing countries, there are several reasons to be-

lieve that returns to education may be higher than current published estimates. First, in our data

wages are positively correlated with education levels in rural China (Table 4). Among individuals

who work for a wage off-farm, wages are significantly higher among people who have attended at

least some post-primary school than among people who have not (row 1). Wages are also higher

among migrants and younger workers who are more educated (rows 2 and 4). For example, mi-

grants who have post-primary education have an average wage of 2.92 yuan per hour, whereas

wages average 2.25 yuan per hour among migrants with a primary education or below.

The method of measuring wages in other rural China studies may further affect estimates of

7

the returns to education. In countries with underdeveloped financial markets, such as China, poorer

people may drop out of school because they cannot finance the earnings they forego to attend

school, while richer people can continue as they wish (Schultz, 1988). As a result, richer people

may systematically have more education. Moreover, after completing education differences in

wealth endowments, which are associated with differences in preferences for leisure and tolerances

for risk, may influence the relatively poor to work more. The poor may have to work either more

hours per day and/or more days per month or year. As a result, estimates of the returns to education

based on daily, monthly or annual earnings could underestimate the true returns to schooling. Since

hourly income is not affected by the choice of how many hours per day or days per month to work,

it is the preferred measure (Schultz, 1988; Card, 1999). None of the previous Mincerian studies in

rural China use hourly wages.

Methodological issues related to the use of unrepresentative samples may further obscure

the measured relationship between wages and completed education. Several papers only consider

workers in one industry or sector of the economy (e.g. Meng, 1996; Gregory and Meng, 1995; Ho

et al., 2001). If that sector had lower rates of return than other sectors, perhaps for institutional

reasons, estimates drawn from such studies would not be representative. Becker (1964) warns that

estimates of educational returns will be low if particular groups of workers are singled out for esti-

mation, because the effect of selection into that portion of the off-farm labor force is ignored. Only

one paper in the rural China literature does a standard Heckman correction for sample selectivity

bias (Zhang et al., 2002).

Finally, previous studies have not considered several other basic empirical issues. For exam-

ple, the returns to education in rural China have not addressed criticisms of the Mincer method that

are well known in the literature (e.g. Schultz, 1988; Strauss and Thomas, 1995; Card, 1999). For

example, none of the Mincerian studies controlled for potential nonlinearities in returns.5 Several

recent studies based on urban China have found highly non-linear returns to education (Heckman

and Li, 2004; Giles et al., 2006). In rural China, if there are higher returns to post-primary ed-

8

ucation than primary education, and previous studies have missed better educated workers, their

estimates would have been biased downward. Since other authors have found a positive correlation

between education and migration (e.g. Zhao, 1999; Rozelle et al., 1999), if other studies missed

or excluded migrants, their estimated returns might be biased downward. In addition, previous

studies have not tried to control for biases that might be created by omitting variables that might be

systematically correlated with schooling levels or returns to education, such as individual ability

or school quality measures.

3 Modeling the Effect of Schooling on Income

While it may not be surprising that the current literature has not found rates of return fit the patterns

outside of China or the patterns that appear in the data, ex ante we do not know what factors

cause such low estimates. To address these shortcomings, we produce new estimates that use

methods that will attempt to address such criticisms. Therefore, in this section, we first present

the model that is most commonly used to estimate the returns to education (Mincer, 1974). The

model is subject to several criticisms, which can be categorized as either problems that result from

measurement issues or omitted variable biases. After summarizing the model, we will discuss

econometric strategy to address the issues.

3.1 The Mincer Model

A simple model relating years of completed schooling to lifetime earnings was proposed by Mincer

(1974). The model assumes, initially, that individuals will work over a lifespan of W years, which

is fixed, are risk neutral, and their human capital stock will not change after they finish school.

Furthermore, the income that an individual will obtain is fixed and related directly to the amount

of schooling they obtain. In this model the private cost of schooling is the cost of delaying the

9

income stream another year, which implies the relationship:

ln(Ys) = ln(Y0) + rS (1)

where Ys is the income after schooling, Y0 is the base income with no schooling, and S is the years

of schooling completed. In our analysis, we use hourly wages as the income measure.

Studies of the returns to education modify equation (1) to account for post-schooling human

capital accumulation, through experience, and sometimes other effects on wages or income (see

Psacharopoulos, 1985; 1994, for surveys). If experience, E, and experience squared are used to

control for post-school human capital accumulation, the return to schooling, r, for individual i can

be determined by estimating:

ln(Ys,i) = ln(Y0) + rSi + β1Ei + β2E2i + εi (2)

Equation (2) is typically referred to as the Mincer “earnings function” method of determining the

returns to schooling, and literally hundreds of studies have estimated the above equation in various

settings. In all specifications, we modify equation (2) by adding a variable that indicates whether

or not an individual has received formal training for a profession or trade (including participation

as a formal apprentice) to account for any wage premium paid for such training, a dummy variable

for gender than controls for the possibility that men and women are paid on different scales, and

provincial dummy variables to account for regional differences in base wage rates. Since marital

status has also been found to affect wage rates in developed countries (e.g. Goldin, 1990), in

most specifications we include a dummy variable for whether or not individuals in our sample are

married.

3.2 Further Adjustments to the Mincer Model

If we simply estimate equation (2) using OLS, we run into several potential biases, and therefore

we adjust our estimation framework to account for them.6 Following Heckman (1974), workers

10

will not enter the workforce if their shadow or reservation wage (Y ∗) is higher than the wage

offered. In rural China, people do not only make a choice between off-farm labor and leisure; they

can also work in farming, raising livestock, and in self-employment. Their individual reservation

wage is therefore determined by other opportunities and/or tradeoffs between labor and leisure. If

that reservation wage is higher than the wage offered for an individual’s skill set, that person will

not enter off-farm labor markets. From the perspective of the analyst, we only observe wages for

individuals who are offered a wage Ys,i > Y ∗. If we do not correct for this selectivity bias, the

estimated return to education in off-farm work may be biased downward.

To account for this potential bias, we first estimate a probit for all individuals in our sample,

where the dependent variable is one if the individual works off-farm for a wage and zero otherwise.

Then, using the results of from the probit estimation we compute an inverse Mills ratio that corrects

for possible truncation of the dependent variable in estimation of equation (2). In order to identify

the probit equation, we include the logarithm of the household’s asset holdings in 1999, the number

of males and females of working age in the household, and the land endowment granted to the

household by the village. None of these variables should affect the hourly wage of a specific

worker, except through their decision about whether or not to participate in off-farm wage labor.

These instruments are should have an effect on an individual’s reservation wages, by affecting the

returns to labor on the farm, in self-employment, and in leisure, but none of them should affect the

wage offered to an individual.7

Equation (2) also implicitly assumes that the returns to education are the same for each

additional year, even when individuals change schools. The returns could be non-linear; that is,

the returns to different levels of schooling are different. To account for variation in the returns to

different levels of schooling, we modify our model to estimate the return to an additional year of

primary school and a separate estimate for an additional year of post-primary school. We modify

equation (2) to estimate different coefficients for the return of primary schooling, rP , and the return

11

to post-primary schooling, rM :

ln(Ys,i) = ln(Y0) + rP Pi + rMMi + β1Ei + β2E2i + εiv (3)

where Pi represents the years of primary school completed by individual i, and Mi represents the

years of post-primary school. If an individual left school before completing primary school, then

Mi is equal to zero.

Additionally, the Mincer model has been criticized for potentially omitting variables cor-

related with both wage levels and schooling, therefore implying bias in the estimated return to

schooling. Therefore we check the robustness of our results to the inclusion of variables that proxy

for an individual’s ability, as many authors have pointed out that if people with more ability to earn

money tend to go to school longer, then the estimated return to schooling will be biased upwards

(e.g. Boissiere, Knight, and Sabot, 1985). Following, among others, Heckman and Hotz (1986),

we attempt to control for individual ability using the educational attainment of the individual’s

mother and father as proxies.

4 Results

To generate estimates of the rates of return to education, we estimated equation (2) while control-

ling for sample selectivity bias (Table 5). In general, the estimator performed well. The Wald test

statistic for the hypothesis that all coefficients were zero was strongly rejected, and OLS versions

of the equations had R2 statistics that range from 0.15 to 0.20. Moreover, most coefficients are of

the expected sign and statistically significant. For example, in almost all reported specifications

we find the expected concave relationship between wages and years of experience. Furthermore,

as described by other authors who have examined the determinants of participation in the off farm

labor market (e.g. Zhao, 1997), we find that education has a strong effect on selection into the

off-farm labor force. At the mean level of education in the sample, given another year of education

12

a person becomes 1.1 to 1.3 percent more likely to find an off-farm job (row 1, columns 1, 3, and

5).

In addition to helping people find off-farm jobs, education has a positive, statistically signif-

icant effect on hourly wages. Depending upon the specification, we find that the average return

to a year of education was between 6.3 an 6.5 percent (row 1, columns 2, 4, and 6). In model 1

(columns 1 and 2), we estimate the average return to education to be 6.3 percent. In model 2, we

include marital status to control for potentially higher wages among married individuals (Goldin,

1990), and find that marital status seems to have a negative effect on participation, and no effect on

wages. In model 3, we test for ability bias by including the educational attainment levels for each

individual’s parents. These variables affect neither off-farm participation nor the wage rate, and so

we choose model 2 as our preferred specification. Therefore our preferred estimate for the average

return to education is 6.4 percent (row 1, column 4).

All three of our estimates are higher than other nationally representative samples from rural

areas listed in Table 2 (e.g. Parish et al., 1995; Johnson and Chow, 1997). One possible explanation

for a higher estimate is that we used a different methodological framework than other studies, as we

specified hourly wages as the dependent variable rather than daily (e.g., Yang, 1997; Meng, 1996;

Gregory and Meng, 1995), monthly (Johnson and Chow, 1997), or annual wages (Ho et al., 2001).

Alternatively, the higher estimate could be due to the fact that we used a more representative

sample, containing observations from across China and including both local wage earners (i.e.,

workers that have a job in local manufacturing or service enterprises) and migrants, instead of a

subsample of local wage earners (e.g. Meng, 1996; Ho et al., 2001). Finally, the rate of return

to education could be increasing over time, and our estimate could be relatively high because we

used recently collected data. Despite our use of a nearly nationally representative sample, recently

collected data, and a more theoretically consistent wage measure, our base estimate is still lower

than estimates normally found for other developing countries.

The rates of return to education, however, rise sharply when the analysis focuses on specific

13

age groups within the population. Among individuals aged 35 and under, the average rate of

return to a year of education rises to 10.5 percent (Table 6; row 1, column 4). This estimate is

consistent with the average return found in the rest of Asia. In contrast, the point estimate for

individuals over 35 is 2.9 percent, and the estimate is statistically insignificant (row 1, column 7).

The difference between returns for younger and older workers may indicate that human capital is

becoming increasingly important for younger workers in rural China, the group that makes up the

largest and fastest increasing part of China’s off-farm labor force.

For migrants, the rate of return also exceeds the national average. When examining the entire

sample, we find that migrants earn a higher return to their education than local wage earners (Table

6; row 1, columns 2 and 3). For all local wage earners, the average return to a year of education

is 4.3 percent (row 1, column 1).8 Among all migrants, the rate of return is 7.8 percent (row 1,

column 3). Moreover, when we restrict the sample to individual migrants 35 years old and under,

we find that the rate of return rises to 11.7 percent (row 1, column 6). Meanwhile, estimates of

the returns to education in both local wage earning and migration are statistically insignificant for

individuals over 35. If migrants, and particularly young migrants, are systematically excluded from

samples, the estimated rate of return could be downward biased.9

The returns to education in rural China might also be underestimated if, for example, later

years of schooling lead to larger wage increments than earlier years of schooling. To test whether

this phenomenon occurs in rural China, equation (3) was estimated using the whole sample and

for individuals aged 35 and under (Table 7). Though primary education has a positive, statistically

significant effect on selection into off-farm work in general (row 1, column 1), the return to a

year of primary schooling is only 3.9 percent, and only marginally statistically significant (row 1,

column 2). Though the return to primary schooling off-farm is low, post-primary schooling has a

large, significant effect on the wage received (row 2, columns 2 and 4). Individuals who find jobs

receive a return to an additional year of post-primary schooling of 7.4 percent (row 2; columns 2).10

These findings are consistent with a story that more educated workers are scarce in rural China,

14

and therefore they are receiving a high return to their education level. Using a data set collected in

urban China, Li (2003) finds a similar low return to primary schooling and suggests that the returns

to primary school are low because nearly every child now completes primary school.

4.1 Reconciling with Other Studies on Rural China

In order to reconcile our estimates for the rate of return to education in rural China with the find-

ings of other authors, we perform the following exercise. First, we attempt to determine how much

of the difference between our estimates and previous estimates are due to the difference in the time

the data was collected, henceforth the timing effect. To do so, we use our data set to mimic the

analyses of previous Mincerian studies (Table 2).11 We follow each author’s wage definition and

econometric specification as closely as possible, and then re-estimate the Mincer model. Since

none of the previous authors did so, in this step we do not correct for selectivity. Where appropri-

ate, we also restrict our sample either geographically or to a certain portion of the off-farm labor

force. The difference between original published estimates and the estimates using our data and

the specifications used by previous authors is one way to determine whether the return to education

has changed over time.

Second, we attempt to find out how much of the differences in rates of return are due to

methodological considerations. We can identify two effects. First, we primarily set out to deter-

mine how much of the difference is due to the mismeasured wage, controlling for sample selectiv-

ity. After using our data to mimic the original analyses, we make corrections in the wage definition

and control for selectivity for all five studies, and re-estimate each model. From this part of the

exercise, we can determine how much of the difference in the results are from the definition of the

dependent variable (and selectivity). As two studies focused solely on the TVE sector, yielding an

incomplete sample of the rural labor force (Meng, 1996; Gregory and Meng, 1995), here we use

their specification and estimate returns using the entire sample, including migrants.

15

Our experiments show that the timing effect accounts for a portion of the difference between

our estimates and the lower estimates found in the literature (Table 8, columns 1 and 2). When

we compare the original rates of return from the literature (that use data collected in years varying

from the late 1980s to the mid-1990s) with the estimates that use our data (from 2000) but their

specifications and wage measurements, rates of return rise in four of the five cases when we use

our data (column 2). On average, the rates of return rise from 2.3 to 3.8 percent, an increase of 1.5

percent.12 While some of the gap that we observe between the rural China literature and the rest

of the world is due to a timing effect, which could be linked to the increasing emergence of labor

markets, the rise in returns is modest. Furthermore, China’s rates are still far below those in the

rest of the world.

The methodological part of the experiment accounts for an even greater part of the gap (Table

8). After replacing each study’s wage definition with the hourly wage and controlling for selectiv-

ity, we find that the rates of return return rise for all studies (columns 2 and 3). In the first three

studies, we only test the mismeasured wage effect, and find the largest increase using the specifi-

cation of Johnson and Chow (1997). For their study, the rates rise from 3.0 to 6.5 percent (row 1).

For Parish et al. (1995) and Yang (1997), the increases are more modest, but still increase by 0.7

and 1.8 percent, respectively (rows 2 and 3). Since the final two studies only examined local wage

earners, we control for both the mismeasured wage and for an unrepresentative sample (Meng,

1996; Gregory and Meng, 1995). While the mismeasured wage effect is found to be positive, the

return only increases by 1.2 percent (4 percent versus 2.8 percent– not shown). However, when we

include migrants when estimating the returns using their specification, the returns rise even further

to 6.1 percent (column 3, rows 4 and 5). In total, the methodological adjustments raise the rates of

return, on average, from 2.7 to 6.1 percent (row 6), accounting for a larger increase than the timing

effect.

Although the timing effect and the two effects associated with methodology help close the

gap between the China literature and the rest of the world, part of the gap still remains. Some of the

16

gap may be due to the nonlinearity of returns. If returns to post-primary education are inherently

higher in rural China, and educational attainment levels are low, part of the gap may be caused

by the fact that a smaller portion of its population is receiving the type of education that creates

higher returns. Besides finding this effect in the previous section (Table 7), we also find it using

the specifications of other authors. After making the corrections for wage definitions and sample

selectivity bias, we find the returns to post-primary school are systematically higher than primary

school in all of the studies, and range between 6.7 and 9.0 percent. The returns to a year of post-

primary school are between 2.3 and 4.9 percent higher than the returns to primary school. As a

result, by not considering the possibility of a convex relationship between wages and education,

all previous studies may have further underestimated the returns to post-primary schooling.

Beyond methodological differences, part of the difference between returns in China and the

rest of the world may also have to do with labor force demographics. Because of birth rates that

declined earlier than in other parts of the developing world, China’s rural labor force is relatively

older than that of other developing countries (e.g. Tu, 2000). To provide a better comparison of

China’s labor force to that of the rest of the world, we concentrate on the rate of return to education

among the young.13 If we revisit the illustration using the data from Psacharopoulos (1994), we see

that the timing of our study, methodology, and the evolution of the labor force account for almost

all of the gap in rates of return and attainment. Correcting for timing and methodology, we boost

rates of return from around 4 to a rate of return of 6.4 percent at an average level of schooling

attainment of around 6 years (Figure 2). When we restrict the sample to individuals under 35 years

old, the rate of return increases to 10.5 percent, at an average schooling level of 7.5 years. When

we plot this point on Figure 2, it is precisely in line with the rates of return and attainment rates

found in the rest of the world.

17

5 Conclusions

In this paper, we have used a nearly nationally representative sample of workers in rural China to

estimate the returns to education in off-farm work. Making corrections for selectivity into off-farm

work and using the hourly wage rate, we find that across all individuals with off-farm jobs in our

sample, the mean return to a year of education is 6.4 percent. The estimated return is even higher

among all migrants in the sample, at 7.8 percent. These estimates are higher than most previously

reported estimates of the return to education in China in the economics literature.

When the sample is restricted to individuals who are aged 35 and under, we find that the

returns to education are 10.5 percent, which equals the average return found in Asia (Psacharopou-

los, 1994). Among migrants, the return to a year of education is even higher, around 11.7 percent.

Furthermore, the returns to education exhibit some convexity, which is consistent with findings of

other authors who have looked for nonlinearities in the returns to education in other developing

countries (e.g. van der Gaag and Vijvenberg, 1989; Strauss and Thomas, 1995; Duflo, 2001) and

in urban China (Li, 2001). The estimates are also robust to controls for individual ability and for

school quality, which may have biased the returns downward.

Our study shows the importance of using reasonably sound methodology on data sets that

are representative of the entire wage earning labor force. When we replicate the methods followed

by other authors to estimate returns to education using these data, we find that their methodology is

associated with lower estimates of the returns to education. The results of this exercise emphasize

the need to use hourly wages, rather than daily or monthly wages, to control for the amount a

person chooses to work, as highly educated people in developing countries tend to choose to work

less. They also show the importance of choosing a representative sample when studying the returns

to education; we find that returns to education tend to be lower when only local labor is studied, as

the returns to education in China are higher for migrants.

Finally, our estimates show that increasing educational availability in rural areas may be

18

a good policy instrument for increasing rural incomes, particularly with respect to post-primary

school. This paper finds that the returns to education among younger workers are at approximately

the same level as returns to schooling in the rest of Asia, and in developing countries in general.

Given these high private returns, China’s government would do well do make rural education a top

priority. As the majority of children in China today are being raised in rural areas, much of China’s

future labor force would be made better off by supporting their education.

19

Notes

1More recently, several authors have estimated much higher rates of return to education in

urban areas (Zhang and Zhao, 2005; Heckman and Li, 2003).

2The provinces included in the survey were Hebei, Shaanxi, Liaoning, Zhejiang, Sichuan, and

Hubei. Since 10 villages were sampled in each province, the sample is almost evenly split between

the six provinces. Descriptive statistics for all variables included in the regressions in the paper are

in Appendix Table 1.

3Meta-analyses of studies of the returns to education average estimates over samples with dif-

fering levels of representativeness as well as over different periods of time. Some of the studies in-

cluded in the averages presented here may share the methodological shortcomings of the literature

on the returns to education in rural China. To the extent that these averages share these shortcom-

ings, we would argue that they represent estimates that are biased slightly downward. Because they

are averaged across many different studies, they should be considered crude estimates, although

representative of the pattern.

4We define the “standard Mincerian method” as regressing the years of education, years of

experience, and years of experience quared on the logarithm of a wage variable while using a

sample of individuals. Seven studies not included in Table 2 calculate the returns to education

using other methods. Zhang et al. (2002) uses a modified Mincer model to estimate the returns to

education in Jiangsu over time; Zhao (1997) and Hare (1999) use other methods to estimate the

returns to education for migrants. Zhang et al. (2002) find high returns at the mean education level

of their sample in 1996 (9%), but find no returns in 1988 and 1992. Zhao and Hare find returns

of 2.6% on average. The other four studies– Benjamin et al. (2000), Li and Zhang (1998), Yang

and An (2002), and Zhao (1999)– all use household level data and find returns of four percent on

average, though Benjamin et al. (2000) is 6.6 percent.

20

5Only two previous studies attempt to explicitly control for a potential nonlinear relationship

between wages and completed schooling by including an “education squared” variable (Zhang et

al., 2002; Yang and An, 2002).

6Although we deal with several of the criticisms of the Mincer model in this paper, it has

further criticisms that we will not discuss here. For example, it does not take into account regional

labor supply and demand relationships that may affect the return to education in one place but not

another (Heckman et al., 1996).

7All of the results in the paper are robust to the exclusion of specific instruments from the

selection equation as well as their inclusion in the wage determination equation.

8We use an identical identification strategy when we limit the sample to local wage earners or

migrants; for example, if we are determining the effect of education on wages paid to migrants, the

local wage they would be offered is captured by the reservation wage. Since different instruments

appear to have different effects on different types of work, we believe the instruments identify the

varying reservation wages (Appendix Table 2). In the full sample and in the subsamples, Hausman

tests reject the hypothesis that the coefficients on the instruments in the local wage earning and

migrants equations are the same.

9This finding is consistent with Taylor and Yunez-Naude (2000), who find that migration has

the highest rate of return among household activities in Mexico.

10Among younger workers, the returns are also convex. Although the point estimate of the return

to primary schooling is somewhat larger, at 6.0 percent, it is only statistically significant at the 10

percent level. However, conditional on holding a job, they receive a return of 11.5 percent to each

year of post-primary schooling.

11We did not replicate the Ho et al. (2001) study, as it covered specific TVEs in Shandong and

Jiangsu provinces, which are not included in our sample.

21

12To make this calculation, we take the average of men (1.8%) and women (4.3%) from the

Parish et al. (1995) study (row 2).

13Furthermore, the education of those between 35 and 50 years old in China may have been

severely affected by the Cultural Revolution. Therefore, it is probably inappropriate to consider a

year of education during the Cultural Revolution and after the Cultural Revolution as the same.

22

References

Alderman, H., Jere R. Behrman, D.R. Ross, and R. Sabot, 1996, The Returns to Endogenous Hu-

man Capital in Pakistan’s Rural Wage Labour Market, Oxford Bulletin of Economics and Statistics

58(1), 29-55.

Appleton, Simon, John Knight, Lina Song, and Qingjie Xia, 2002, Towards a Competitive Labour

Market? Urban Workers, Rural Migrants, Redundancies and Hardships in China, Institute for

Contemporary China Studies Working Paper. Nottingham: University of Nottingham.

Becker, Gary, 1964, Human Capital (Columbia University Press, New York).

Behrman, Jere, and Nancy Birdsall, 1983, The Quality of Schooling: Quantity Alone is Mislead-

ing, American Economic Review 73(5), 928-946.

Benjamin, Dwayne, Loren Brandt, Paul Glewwe, and Guo Li, 2000, Markets, Human Capital, and

Inequality: Evidence from Rural China, Working Paper, Department of Economics, University of

Toronto.

Boissiere, M., J.B. Knight and R. Sabot, 1985, Earnings, Schooling, Ability, and Cognitive Skills,

American Economic Review 75(5), 1016-1030.

Brandt, Loren, Jikun Huang, Guo Li, and Scott Rozelle, 2002 (Jan.), Land Rights in China: Fact,

Fiction, and Issues, China Journal 47, 67-97.

Brown, Phil and Albert Park, 2002, Education and Poverty: The Effects of Credit Constraints,

Intra-Household Decision Making, and School Quality on Educational Attainment in China’s Poor

Areas, Economics of Education Review, forthcoming.

23

Card, David, 1999, The Causal Effect of Education on Earnings, Handbook of Labor Economics

(Elsevier Science Publishers, Amsterdam), Orley Ashenfelter and David Card, eds., Volume 3A,

Chapter 20, pp. 1801-1863.

Chen, Hongyi, and Scott Rozelle, 1999, Leaders, Managers, and the Organization of Township and

Village Enterprises in China, Journal of Development Economics 60(2), 529-557.

Chou, E.C., and L.J. Lau, 1987, Farmer ability and farm productivity: A study of the farm house-

holds in the Chiangmai Valley, Thailand, 1972-1978, Discussion Paper 62, Education and Training

Department, The World Bank, Washington, DC.

Davidson, Russell, and James G. MacKinnon, 1997, Estimation and Inference in Econometrics

(Oxford University Press, Oxford).

Duflo, Esther, 2001, Schooling and Labor Market Consequences of School Construction in Indone-

sia: Evidence from an Unusual Policy Experiment, American Economic Review 91(4), 795-813.

Gregory, R.G., and Xin Meng, 1995, Wage Determination and Occupational Attainment in the

Rural Industrial Sector of China, Journal of Comparative Economics 21, 353-374.

Hare, Denise, 1999, ‘Push’ versus ‘Pull’ Factors in Migration Outflows and Returns: Determinants

of Migration Status and Spell Duration among China’s Rural Population, Journal of Development

Studies 35(3): 45-72.

Heckman, James, 1974, Shadow Prices, Market Wages, and Labor Supply, Econometrica 42, 679-

94.

24

Heckman, James and V.J. Hotz, 1986, An investigation of the labor market earnings of Panamanian

males: Evaluating the Sources of Inequality, Journal of Human Resources 21(4), 507-542.

Heckman, James, Anne Layne-Farrar, and Petra Todd, 1996, Human Capital Pricing Equations

with an Application to Estimating the Effect of Schooling Quality on Earnings, Review of Eco-

nomics and Statistics 78(4), 562-610.

Heckman, James, and Xuesong Li, 2003, Selection Bias, Comparative Advantage, and Heteroge-

neous Returns to Education: Evidence from China in 2000, National Bureau of Economic Research

Working Paper 9877.

Ho, Samuel, Xiaoyuan Dong, Paul Bowles, and F. MacPhail, 2001, Privatization and Enterprise

Wage Structures During Transition: Evidence from China’s Rural Industries, Economics of Tran-

sition, forthcoming.

Jin, Hehui, and Yingyi Qian, 1998, Public versus Private Ownership of Firms: Evidence from

Rural China, Quarterly Journal of Economics, 773-808.

Johnson, Emily N., and Gregory C. Chow, 1997, Rates of Return to Schooling in China, Pacific

Economic Review 2(2), 101-113.

Li, Haizheng, 2003, Economic Transition and Returns to Education in China, Economics of Edu-

cation Review 22(3), pp. 317-328.

Li, Tianyou, and Junsen Zhang, 1998, Returns to education under collective and household farming

in China, Journal of Development Economics 56, 307-335.

Maurer-Fazio, Margaret, 1999, Earnings and education in China’s transition to a market economy:

Survey evidence from 1989 and 1992, China Economic Review 10, 17-40.

25

Meng, Xin, 1996, An Examination of wage determination in China’s rural industrial sector, Applied

Economics 28, 715-724.

Meng, Xin, and Junsen Zhang, 2001, The Two-Tiered Labor Market in Urban China: Occupational

Segregation and Wage Differentials between Urban Residents and Rural Migrants in Shanghai,

Journal of Comparative Economics 29, 485-504.

Mincer, Jacob, 1974, Schooling, experience, and earnings (Columbia University Press, New York).

Naughton, Barry, 1995, Growing Out of the Plan: Chinese Economic Reform 1978-1993 (Cam-

bridge University Press, Cambridge, UK).

Nyberg, Albert, and Scott D. Rozelle, 1999, Accelerating China’s Rural Transformation (World

Bank, Washington D.C.).

Parish, William, Xiaoye Zhe and Fang Li, 1995, Non-farm work and marketization of the Chinese

countryside, The China Quarterly 143, 697-730.

Psacharopoulos, George, 1985, Returns to education: a further international update and compari-

son, The Journal of Human Resources, 20(4), 583-597.

Psacharopoulos, George, 1994, Returns to Investment in Education: A Global Update, World De-

velopment 22(9), 1325-1343.

Psacharopoulos, George, and E. Velez, 1992, Schooling, ability, and earnings in Colombia, 1988,

Economic Development and Cultural Change 40(3), 629-643.

26

Rozelle, Scott, Li Guo, Minggao Shen, Amelia Hughart, and John Giles, 1999, “Leaving China’s

farms: Survey results of new paths and remaining hurdles to rural migration,” The China Quarterly

158, 367-393.

Schultz, T. Paul, 1988, Education Investments and Returns, in Hollis Chenery and T.N. Srinivisan,

eds., Handbook of Development Economics, Volume I (Elsevier Science Publishers: Amsterdam),

544-630.

Strauss, John, and Duncan Thomas, 1995, Human Resources: Empirical Modeling of Household

and Family Decisions, in Jere Behrman and T.N. Srinivisan, eds., Handbook of Development Eco-

nomics, Volume 3A (Elsveier Science Publishers: Amsterdam), 1885-2023.

Tsang, Mun C., 1996, Financial Reform of Basic Education in China, Economics of Education

Review 15(4), 423-444.

Tu, Ping, 2000, Trends and Regional Differentials in Fertility Transition, chapter 3 in: Xizhe

Pen and Zhigang Guo, eds., The Changing Population of China (Blackwell Publishers: Malden,

Massachusetts), 22-33.

van der Gaag, Jacques, and Wim Vijverberg, 1989, Wage Determinants in Cote d’Ivoire: Ex-

perience, Credentials, and Human Capital, Economic Development and Cultural Change 37(2),

371-81.

West, Lorraine A., 1997, Provision of Public Services in rural PRC, chapter 6 in Financing Local

Government in the People’s Republic of China, Christine Wong, ed. (Hong Kong, Asian Develop-

ment Bank), pp. 213-282.

World Bank, 2002, World Development Indicators CD-ROM.

27

Yang, Dennis T., 1997, Education and Off-Farm Work, Economic Development and Cultural

Change 45, 613-632.

Yang, Dennis T., and Mark An, 2002, Human Capital, Entrepreneurship, and Farm Household

Earnings, Journal of Development Economics 68(1), 65-88.

Zhang, Junsen, and Yaohui Zhao, 2002, Economic Returns to Schooling in Urban China, 1988-

1999, working paper, Beijing University.

Zhang, Linxiu, Jikun Huang, and Scott Rozelle, 2002, Employment, Emerging Labor Markets, and

the Role of Education in Rural China, China Economic Review 13(2-3), 313-328.

Zhao, Yaohui, 1997, Labor Migration and Returns to Rural Education in China, American Journal

of Agricultural Economics 79(4), 1278-87.

Zhao, Yaohui, 1999, Labor migration and earnings differences: The case of rural China, Economic

Development and Cultural Change 47(4), 767-82.

ZGTJNJ (Zhongguo Tongji Nianjian– China Statistical Year Book), 2001, (Beijing, China: China

Statistical Press).

28

Table 1: Share of Labor Force Participating in Different Activities, Rural China

Sample Farming Local Wage Self-Employment MigrationLabor

NBS (2001) 0.670 0.330*Our Sample, All Individuals 0.762 0.133 0.162 0.170Our Sample, No Migrants 0.847 0.161 0.195

Notes: NBS statistics do not differentiate the self-employed from workers earning wages locally. In oursample, workers can participate in both farming and one off-farm activity, and therefore totals sum to morethan 1.Source: Authors’ survey.

29

Table 2: Returns to Education in Rural China: Other Studies using the Mincer Method

Dependent Years Data Analytical Returns toAuthors(Date) Variable Cover Approach Education

Johnson and Chow (1997) Yuan/month 1988 Individuals 4.0Parish et al. (1995) Yuan/year 1993 Individuals 1.8-4.3Yang (1997) Yuan/day 1990 Individuals 2.3Meng (1996) Yuan/day 1986-7 TVE workers 0.7-1.1Gregory and Meng (1995) Yuan/day 1986-7 TVE workers 0.7-1.1Ho et al. (2001) Yuan/year 1998 TVE workers 3.2-5.1

30

Table 3: Hourly Wages, by Type of Work, Cohort, and School AttendanceAttended Any

Post-Primary School?N no yes

All individuals 1022 2.48 2.99**(1.95) (2.66)

Migrants 571 2.25 2.92**(1.63) (2.38)

Individuals under 35 672 2.43 2.77*(2.19) (2.37)

Migrants under 35 479 2.16 2.83**(1.63) (2.38)

Notes: Standard deviations in parentheses. **- indicates difference between means significant at the 5%level; *- indicates difference between means significant at the 10% level.

31

Table 4: Effects of Education and Experience of Off-Farm WagesModel 1 Model 2 Model 3

Selection Wage Selection Wage Selection WageEquation Equation Equation Equation Equation Equation

Years of 0.011 0.063 0.013 0.064 0.013 0.065Education (3.28)** (7.52)** (4.03)** (7.37)** (4.02)** (7.30)**Years of -0.011 0.025 0.001 0.031 0.002 0.035Experience (6.00)** (4.70)** (0.22) (4.55)** (0.57) (4.83)**Experience, 0.003 -0.048 -0.013 -0.056 -0.015 -0.060Squared (/100) (0.78) (4.81)** (3.13)** (4.29)** (3.24)** (4.52)**Gender 0.078 0.220 0.198 0.210 0.202 0.221(1=male) (3.60)** (3.53)** (11.44)** (3.48)** (11.50)** (3.51)**Marital Status -0.200 -0.082 -0.221 -0.106(1=married) (6.29)** (0.98) (6.65)** (1.18)Skill Training 0.205 0.130 0.083 0.131 0.088 0.130(1=yes) (11.93)** (2.67)** (3.82)** (2.64)** (3.96)** (2.56)**Father’s Years -0.001 0.011of Education (0.42) (1.60)Mother’s Years of -0.0001 -0.0001Education (0.54) (0.19)InstrumentsLog, Value of Household -0.051 -0.047 -0.048Assets in 1999 (7.20)** (6.69)** (6.61)**Number of Male -0.020 -0.027 -0.026Workers (1.68)* (2.28)** (2.09)**Number of Female 0.067 0.068 0.066Workers (6.20)** (6.31)** (6.01)**Land Endowment -0.005 -0.005 -0.005

(3.40)** (3.30)** (3.40)**Inverse Mills 0.234 0.219 0.232Ratio (1.81)* (1.68)* (1.74)*Number of Obs. 3364 1023 3364 1023 3235 986

Notes: t-statistics are in parentheses. *- indicates significance at the 10 percent level; **- indicatessignificance at the 5 percent level. Provincial fixed effects are included in all equations. Experienceis measured as years since the person left school if they went to school, and age−6 if they did not.All regressions are done using the two-step method proposed by Heckman (1974) and standarderror calculations take the method into account.

32

Tabl

e5:

Ret

urns

toSc

hool

ing,

byC

ohor

tand

Type

ofO

ff-F

arm

Wor

kA

llIn

divi

dual

sIn

divi

dual

s35

and

unde

rIn

divi

dual

sove

r35

Loc

al+

Loc

alW

age

Mig

rant

sL

ocal

+L

ocal

Wag

eM

igra

nts

Loc

al+

Loc

alW

age

Mig

rant

sM

igra

nts

Ear

ning

Mig

rant

sE

arni

ngM

igra

nts

Ear

ning

Yea

rsof

0.06

40.

043

0.07

80.

105

0.08

90.

117

0.02

90.

026

-0.0

10E

duca

tion

(7.3

7)**

(2.9

4)**

(6.5

8)**

(8.2

0)**

(3.1

5)**

(8.0

6)**

(1.9

3)(1

.37)

-(0.

26)

Yea

rsof

0.03

10.

039

0.02

70.

084

0.09

10.

089

0.02

60.

037

-0.0

46E

xper

ienc

e(4

.55)

**(2

.94)

**(2

.54)

**(5

.11)

**(2

.53)

**(4

.71)

**(1

.33)

(1.5

5)(0

.83)

Exp

erie

nce,

-0.0

56-0

.059

-0.0

79-0

.197

-0.1

77-0

.236

-0.0

45-0

.050

-0.0

08Sq

uare

d(/

100)

(4.2

9)**

(2.4

9)**

(3.9

6)**

(3.0

5)**

(1.3

5)(2

.94)

**(1

.60)

(1.4

3)(0

.11)

Gen

der

0.21

00.

176

0.23

90.

116

0.16

10.

126

0.39

00.

193

1.50

6(1

=mal

e)(3

.48)

**(1

.47)

(3.4

4)**

(2.0

7)**

(1.1

6)(2

.03)

**(1

.88)

(0.9

1)(2

.41)

**M

arri

ed?

-0.0

82-0

.016

-0.2

17-0

.252

-0.1

24-0

.331

0.16

40.

106

0.70

8(1

=yes

)(0

.98)

(0.1

3)(1

.78)

(2.3

9)**

(0.7

0)(2

.27)

**(0

.75)

(0.4

0)(1

.24)

Skill

Trai

ning

?0.

131

0.13

90.

169

0.09

00.

032

0.13

00.

232

0.23

80.

457

(1=y

es)

(2.6

4)**

(1.9

4)(2

.38)

**(1

.53)

(0.2

9)(1

.82)

(2.6

9)**

(2.5

2)**

(1.6

8)

Inve

rse

Mill

s0.

219

-0.0

110.

556

0.30

10.

351

0.47

40.

079

-0.1

921.

337

Rat

io(1

.68)

(0.0

5)(2

.99)

**(2

.14)

**(1

.15)

(2.5

4)**

(0.2

7)(0

.54)

(1.6

9)

Num

bero

fObs

.33

6433

6433

6415

1515

1515

1519

0619

0619

06N

otes

:t-

stat

istic

sin

pare

nthe

ses.

**-i

ndic

ates

sign

ifica

nce

atth

e5

perc

ent

leve

l.A

llre

gres

sion

sar

edo

neus

ing

the

two-

step

met

hod

prop

osed

byH

eckm

an(1

974)

and

stan

dard

erro

rca

lcul

atio

nsta

keth

em

etho

din

toac

coun

t,an

dfir

st-s

tage

prob

itsas

soci

ated

with

thes

ere

gres

sion

sar

epr

esen

ted

inA

ppen

dix

Tabl

e??

.

33

Table 6: Differential effects of primary education, and post-primary education on wagesAll Individuals Individuals 35 and under

Explanatory Selection Wage Selection WageVariables Equation Equation Equation EquationYears of 0.016 0.039 0.046 0.060Primary School (2.58)** (1.99)** (2.95)** (1.80)Years of Post- 0.011 0.074 0.016 0.115Primary School (2.29)** (6.55)** (1.85) (7.86)**Years of 0.000 0.033 0.021 0.088Experience (0.10) (4.74)** (2.31)** (5.35)**Experience, -0.013 -0.060 -0.065 -0.229Squared (/100) (2.83)** (4.53)** (1.79) (3.39)**Gender 0.197 0.206 0.172 0.113(1=male) (11.35)** (3.43)** (5.50)** (2.02)**Marital Status -0.200 -0.079 -0.338 -0.230(1=married) (6.26)** (0.95) (8.27)** (2.16)**Skill Training? 0.083 0.131 0.096 0.088(1=yes) (3.80)** (2.65)** (2.96)** (1.51)Log, Value of -0.047 -0.077HH Assets, 1999 (6.69)** (6.28)**Number of -0.027 -0.035Male Workers (2.26)** (1.65)Number of 0.068 0.109Female Workers (6.26)** (5.74)**Land -0.005 -0.006Endowment (mu) (3.30)** (2.22)**

Inverse Mills 0.205 0.288Ratio (1.58) (2.06)**

Number of Obs. 3364 1023 1515 672Notes: t-statistics are in parentheses. **- indicates statistical significance at the 95% level. Provincial fixedeffects are included in all equations. Experience is measured as years since the person left school if theywent to school, and age − 6 if they did not. All regressions are done using the two-step method proposedby Heckman (1974) and standard error calculations take the method into account.

34

Table 7: Returns to Schooling, using alternative specifications

Using 2000 Data SetTheir Their Our

Paper Estimate Method MethodJohnson and Chow (1997) 4.0% 3.0% 6.5%Parish et al. (1995) 1.8-4.3% 5.7% 6.4%Yang (1997) 2.3% 4.9% 6.7%Meng (1996) 1.1% 2.7% 6.1%Gregory and Meng (1995) 1.1% 2.7% 6.1%

Notes: Estimates in column 2 were made by applying the methodology listed in thepaper to the data set described in this paper. Column 3 includes the estimate for thedata in this paper using our method, including only the subsample of workers used intheir papers where appropriate. All estimates are significant at the 95 percent level.

35

Years of Education4 6 8 10 12

4

6

8

10

12

14

Asia

Sub-Saharan Africa

Latin America

Middle East

OECD

Rural China

Notes: The five regional are from Psacharopoulos (1994), Table 4. “Asia” does not include Japan, the“Middle East” includes all non-OECD European countries in Europe and North Africa, and “Latin America”includes all countries in the Caribbean. The point for rural China are based on the average level of educationin the data used in this paper, and the returns to schooling are an average of the papers that use the Mincermethod in Table 2.

Figure 1: Educational Attainment and Returns to Schooling

36

Years of Education4 6 8 10 12

4

6

8

10

12

14

Asia

Sub-Saharan Africa

Latin America

Middle East

OECD

Rural China (old)

Rural China (ours)

Rural China*

Notes: The five regional points are from Psacharopoulos (1994), Figure 2. The point for rural China arebased on the average level of education and the returns to education found for individuals aged 35 and underin this paper. The point “Rural China (ours)” reflects our estimates for the entire sample; the point “RuralChina*” reflects estimates for individuals under 35.

Figure 2: Educational Attainment and Returns to Education, Revisited

37

Appendix Table 1: Summary statisticsOff-Farm Workforce All Others

Variable Mean Std. Dev. Mean Std. Dev.All individualsHourly wage 2.83 2.47 − −Years of Education 7.68 2.97 5.45 3.56Years of Primary School 5.45 1.33 4.25 2.32Years of Post-Prim. School 2.22 2.15 1.19 1.76Years of Experience 17.46 13.27 29.48 15.82Skill Training? 0.30 0.46 0.16 0.36Good grades? (1=yes) 0.25 0.43 0.18 0.38Bad grades? (1=yes) 0.13 0.34 0.11 0.31Father’s Years of Ed. 4.15 3.54 2.87 3.29Mother’s Years of Ed. 2.17 3.08 1.19 2.48Migrant? (1=yes) 0.56 0.50 − −Individuals under 35Hourly wage 2.69 2.33 − −Years of Education 8.13 2.68 7.16 2.94Years of Primary School 5.68 0.94 5.30 1.49Years of Post-Prim. School 2.44 2.17 1.86 1.95Years of Experience 9.49 6.33 12.97 7.02Skill Training? 0.31 0.46 0.22 0.41Good grades? (1=yes) 0.23 0.42 0.17 0.37Bad grades? (1=yes) 0.15 0.35 0.15 0.36Father’s Years of Ed. 5.28 3.36 4.67 3.41Mother’s Years of Ed. 2.96 3.32 2.54 3.21Migrant? (1=yes) 0.72 0.45 − −

Notes: 1022 individuals are off-farm workers and are included in columns 1-2; 2341 individuals are not andare included in columns 3-4. Of the off-farm workers, 672 are 35 or under, and 841 are 35 or under and donot work off-farm. For 128 observations, the father’s education level was unknown, so these observationsare not included.

38

Appendix Table 2: Multinomial Logit Regressions Describing Activity ChoiceAll Individuals Individuals Aged 35 and under

Migration Local Wage Self- Migration Local Wage Self-Employment Employment Employment Employment

Years of 0.038 0.085 -0.008 0.130 0.136 0.137Education (1.34) (3.19)** (0.32) (3.20)** (2.72)** (2.79)**Years of -0.007 0.063 0.030 0.196 0.211 0.351Experience (0.25) (2.59)** (1.27) (3.67)** (2.43)** (5.07)**Experience, -0.135 -0.153 -0.109 -0.715 -0.587 -1.086Squared (/100) (2.16)** (3.36)** (2.54)** (3.48)** (1.91) (4.30)**Gender 1.226 1.574 1.384 0.934 1.506 1.464(1=male) (8.32)** (9.16)** (10.72)** (5.73)** (6.45)** (8.27)**Married? -1.163 -0.353 0.404 -1.812 -0.921 -0.245(1=yes) (5.32)** (1.63) (1.74) (6.38)** (2.75)** (0.72)Skill Training? 0.988 0.716 1.051 0.805 0.686 0.886(1=yes) (6.67)** (4.45)** (7.36)** (5.16)** (2.94)** (4.04)**Log, Household -0.152 -0.006 0.546 -0.203 -0.061 0.585Assets, 1999 (2.54)** (0.09) (9.18)** (3.32)** (0.64) (5.86)**Number of -0.080 -0.371 -0.138 -0.060 -0.524 0.025Male Workers (0.78) (3.61)** (1.11) (0.51) (2.94)** (0.20)Number of 0.378 0.204 -0.152 0.467 0.466 -0.129Female Workers (4.77)** (2.63)** (1.51) (4.96)** (3.77)** (0.85)Land -0.034 -0.035 -0.034 -0.026 -0.027 -0.010Endowment (mu) (2.91)** (2.31)** (2.97)** (2.22)** (1.42) (0.75)Regression StatisticsNumber of Obs. 3364 1515Pseudo Log-Likelihood -3151.9 -1599.4Pseudo R2 0.22 0.179

Notes: Asymptotic z-statistics in parentheses are based on standard errors clustered at the village level. **- indicatessignificance at the 5 percent level. Columns 1-3 represent the first regression; columns 4-6 represent the secondregression. Both regressions include provincial dummy variables.

39

App

endi

xTa

ble

3:D

eter

min

ants

ofO

ff-F

arm

Lab

orPa

rtic

ipat

ion,

byC

ohor

tand

Type

ofO

ff-F

arm

Wor

kA

llIn

divi

dual

sIn

divi

dual

s35

and

unde

rIn

divi

dual

sove

r35

Loc

al+

Loc

alW

age

Mig

rant

sL

ocal

+L

ocal

Wag

eM

igra

nts

Loc

al+

Loc

alW

age

Mig

rant

sM

igra

nts

Ear

ning

Mig

rant

sE

arni

ngM

igra

nts

Ear

ning

Yea

rsof

0.01

30.

009

0.00

20.

024

0.00

70.

015

0.00

80.

009

-0.0

01E

duca

tion

(4.0

3)**

(4.1

2)**

(0.8

2)(3

.57)

**(1

.79)

(2.4

9)**

(2.2

7)**

(2.9

4)**

(0.7

7)Y

ears

of0.

001

0.00

7-0

.005

0.02

60.

009

0.01

6-0

.002

0.00

4-0

.003

Exp

erie

nce

(0.2

2)(4

.09)

**(2

.56)

**(2

.93)

**(1

.58)

(2.0

3)**

(0.4

0)(1

.07)

(1.7

1)E

xper

ienc

e,-0

.013

-0.0

13-0

.003

-0.0

88-0

.020

-0.0

73-0

.005

-0.0

100.

002

Squa

red

(/10

0)(3

.13)

**(4

.55)

**(0

.88)

(2.6

2)**

(0.9

8)(2

.35)

**(0

.86)

(1.8

6)(0

.74)

Gen

der

0.19

80.

108

0.06

70.

173

0.09

30.

069

0.19

80.

130

0.05

2(1

=mal

e)(1

1.44

)**

(9.0

0)**

(5.6

3)**

(5.5

3)**

(4.9

3)**

(2.3

3)**

(11.

11)*

*(8

.24)

**(5

.98)

**M

arri

ed?

-0.2

00-0

.007

-0.1

39-0

.335

-0.0

13-0

.311

0.00

0-0

.007

-0.0

01(1

=yes

)(6

.29)

**(0

.35)

(5.5

0)**

(8.2

1)**

(0.4

8)(8

.22)

**(0

.01)

(0.1

7)-(

0.04

)Sk

illTr

aini

ng?

0.08

30.

015

0.05

50.

098

0.01

60.

081

0.04

00.

006

0.02

0(1

=yes

)(3

.82)

**(1

.04)

(3.5

7)**

(3.0

2)**

(0.8

2)(2

.64)

**(1

.67)

(0.2

9)(1

.83)

Log

,Val

ueof

-0.0

47-0

.011

-0.0

26-0

.076

-0.0

10-0

.062

-0.0

21-0

.011

-0.0

06H

HA

sset

s,19

99(6

.69)

**(2

.35)

**(5

.70)

**(6

.22)

**(1

.42)

(5.6

3)**

(3.0

3)**

(1.7

8)(2

.16)

**N

umbe

rofM

ale

-0.0

27-0

.032

0.00

3-0

.039

-0.0

470.

013

-0.0

09-0

.014

0.00

3W

orke

rs(2

.27)

**(3

.77)

**(0

.35)

(1.8

5)(3

.50)

**(0

.66)

(0.6

6)(1

.21)

(0.6

9)N

umbe

rofF

emal

e0.

068

0.01

90.

035

0.10

90.

028

0.07

40.

021

0.01

00.

007

Wor

kers

(6.3

0)**

(2.6

1)**

(5.2

3)**

(5.7

6)**

(2.5

2)**

(4.5

0)**

(1.7

5)(0

.94)

(1.4

8)L

and

-0.0

05-0

.002

-0.0

02-0

.005

-0.0

01-0

.004

-0.0

05-0

.003

-0.0

01E

ndow

men

t(m

u)(3

.30)

**(1

.97)

**(2

.01)

**(2

.16)

**(0

.73)

(1.7

4)(2

.73)

**(2

.36)

**(1

.05)

Not

es:A

sym

ptot

icz

-sta

tistic

sin

pare

nthe

ses.

**-i

ndic

ates

sign

ifica

nce

atth

e5

perc

entl

evel

.Cor

resp

ondi

ngse

cond

stag

ere

gres

sion

sar

ein

Tabl

e6.

40